Professional Documents
Culture Documents
This review considers computational psychiatry from a particular viewpoint: namely, a commitment to explaining psychopathology
in terms of pathophysiology. It rests on the notion of a generative model as underwriting (i) sentient processing in the brain, and (ii)
the scientific process in psychiatry. The story starts with a view of the brain—from cognitive and computational neuroscience—as
an organ of inference and prediction. This offers a formal description of neuronal message passing, distributed processing and
belief propagation in neuronal networks; and how certain kinds of dysconnection lead to aberrant belief updating and false
inference. The dysconnections in question can be read as a pernicious synaptopathy that fits comfortably with formal notions of
how we—or our brains—encode uncertainty or its complement, precision. It then considers how the ensuing process theories are
tested empirically, with an emphasis on the computational modelling of neuronal circuits and synaptic gain control that mediates
attentional set, active inference, learning and planning. The opportunities afforded by this sort of modelling are considered in light
of in silico experiments; namely, computational neuropsychology, computational phenotyping and the promises of a
computational nosology for psychiatry. The resulting survey of computational approaches is not scholarly or exhaustive. Rather, its
1234567890();,:
aim is to review a theoretical narrative that is emerging across subdisciplines within psychiatry and empirical scales of investigation.
These range from epilepsy research to neurodegenerative disorders; from post-traumatic stress disorder to the management of
chronic pain, from schizophrenia to functional medical symptoms.
Wellcome Centre for Human Neuroimaging, Institute of Neurology, University College London, London WC1N 3AR, UK. ✉email: k.friston@ucl.ac.uk
1
Fig. 1 This figure summarizes hierarchical message passing in predictive coding. Predictive coding is a scheme for generative models
based upon continuous states. In these schemes, neuronal activity encodes expectations about the causes of sensory input that are
continually updated to minimize prediction error. Prediction error is the difference between (ascending) sensory input and (descending)
predictions of that input. This minimization rests upon recurrent neuronal interactions among different levels of a cortical hierarchy. It is
thought that superficial pyramidal cells (red triangles) compare the expectations (at each hierarchical level) with top-down predictions from
deep pyramidal cells (black triangles) of higher levels: see [66] for a review. Left panel: this schematic shows a simple cortical hierarchy with
ascending prediction errors and descending predictions. It includes neuromodulatory gating or gain control (teal) of superficial pyramidal
cells that determines their relative influence on deep pyramidal cells encoding expectations. Right panel: this provides a schematic example in
the visual system: it shows the putative cells of origin of ascending or forward connections that convey prediction errors (red arrows) and
descending or backward connections (black arrows) that furnish predictions. The prediction errors are weighted by their expected precision,
which have been associated with projections from the pulvinar. In this example, the frontal eye fields send predictions to primary visual
cortex, which projects to the lateral geniculate body. However, the frontal eye fields also send proprioceptive predictions to pontine nuclei,
which are passed to the oculomotor system to induce movement, through classical oculomotor reflexes. This means that saccadic eye
movements, for example, realise top-down proprioceptive predictions that are deeply informed by hierarchical processing of data from all
sensory modalities. Similar schemes can be elaborated for discrete state space generative models, responsible for planning and motor control
—or allostasis and autonomic function: see [54] for equivalent schematics and [70] for mixtures of continuous and discrete models.
hierarchy; in which prediction errors ascend from lower to higher an intimate relationship between attention and the modulation of
levels to drive changes in neuronal populations encoding states of synaptic efficacy by classical neuromodulators and nonlinear
affairs in the world. These populations (e.g., deep pyramidal cells postsynaptic responses responsible for mediating the exchange
in lower layers of the cortex) then supply a counter stream of between fast-spiking inhibitory interneurons and (superficial)
descending predictions that resolve or cancel prediction errors at pyramidal cells [78, 79, 81, 86–91]. Please see Table 1 and [92]
lower levels [66, 75, 77]: e.g., by targeting inhibitory interneurons for a fuller discussion of neuromodulators and the representation
that are coupled to the superficial pyramidal cells broadcasting of precision. Figures 2, 3 illustrate the neuronal circuitry implicated
prediction errors [78, 79]. This architecture also plays out under in precision or gain control, with a focus on canonical microcircuits
discrete generative models and has become something of a and inhibitory interneurons, respectively.
workhorse for understanding recurrent message passing in One crucial aspect of this precision engineering is that it
cortical and subcortical hierarchies. underwrites our ability to filter out—or ignore—certain prediction
errors when they are deemed imprecise. A key example is the
attenuation of sensory prediction errors that report the con-
THE IMPORTANCE OF BEING PRECISE sequences of movement [93–98]. If we could not ignore the
Prediction errors (i.e., the gradients that drive belief updating) can proprioceptive and somatosensory afferents—supplying evidence
be regarded as carrying the newsworthy information at any given that we are not moving—then any beliefs about intended or
hierarchical level to the level above. However, this is not the predicted movement would be revised immediately, and we
complete story. Higher levels have to select which prediction would not be able to initiate movement. This mandates a transient
errors to listen to; much in the same way that we select our most suspension of sensory precision during active sensing: c.f.,
trustworthy news channels or sources of information. This saccadic suppression of optic flow signals during saccadic eye
selection rests upon predictions of predictability or precision. movements [99]. It also provides a glimpse of how one might
Predicting or estimating precision is a universal requisite for explain Parkinson’s disease in terms of dopaminergic failures of
making sense of data: from estimating the signal-to-noise ratio in sensory attenuation [100]. More generally, it foregrounds the role
some sensory apparatus, through to estimating standard error of precision in orchestrating the perception action cycles [101]
when making inferences via some Neyman-Pearson statistic. that underwrite active sensing [102]. For example, sensory
Affording certain prediction errors greater precision increases attenuation suggests a particular scheduling of action and
their influence on belief updating and has all the hallmarks of perception: although both work hand-in-hand to resolve predic-
attentional selection [80–85]. Physiologically, this simply entails an tion errors, this resolution may involve a fast (~4 Hz) alternation
increase in the excitability or postsynaptic gain of neuronal between sampling the world (e.g., visual palpation with saccades)
populations broadcasting prediction errors. On this view, there is and belief updating after each sample.
Fig. 2 Predicting precision. This schematic is a more detailed version of Fig. 1 that includes putative laminar-specific connections that are
consistent with the precision-based predictive coding described in [176]. This architecture is based upon [66] and conforms roughly to the
known neuroanatomy and physiology of canonical microcircuits in the visual system and the laminar specificity of extrinsic connections. The
key aspect of this figure is the inclusion of deep pyramidal cells encoding the amplitude of (squared or unsigned) prediction error that inform
posterior expectations about precision in the (matrix cells) of the pulvinar. These cells reciprocate descending projections to modulate the
gain of superficial pyramidal cells in cortex. Forward connections are in red, and descending (backward) connections are in black. First-order
prediction errors are shown as full red lines and second-order (unsigned) prediction errors are shown as broken red lines. This schematic
ignores inhibitory interneurons that mediate some inferences vicariously. See subsequent figure.
In summary, the narrative so far is that psychopathology disruption to the organs of connection (i.e., white matter tracts in
represents false inference or aberrant belief updating, under a the brain), positing a form of disconnection syndrome [104]. A
view of the brain as a statistical organ, generating predictions and complementary perspective was provided by Bleuler’s notion of
revising its (subpersonal Bayesian) beliefs on the basis of disintegration of the psyche [105], cast in more functional terms
prediction errors. Crucially, these predictions are contextualised that we might now read as a failure of synaptic integration. These
with predictions of precision or predictability that instantiate ideas re-emerged with the advent of functional brain imaging in the
attentional or intentional set; allowing the selection of attenuation 1990s, in the form of disconnection hypotheses, drawing analogies
of prediction errors via a process of precision weighting. This with things like metachromatic leukodystrophy [106] and the
precision weighting is nothing more than modulating the gain, dysconnection hypothesis [107] that emphasised dysfunctional
postsynaptic sensitivity or excitability of appropriate neuronal synaptic integration; specifically, neuromodulatory failures of
populations. So, what does this computational architecture offer in synaptic function [34]. The latter formulation of dysfunctional
terms of potential targets for pathology? connectivity is now arguably the prevalent view for most psychiatric
and neurodegenerative conditions, while variants of the sejunction
hypothesis would be apt for cerebrovascular accidents, space
SYNAPTOPATHY, DYSCONNECTIONS AND FALSE INFERENCE occupying lesions and other interruptions to white-matter fasciculi:
The narrative returns here to the 19th-century and early formula- for example, increases in conduction delays due to cerebral small
tions of psychiatric disorders such as dementia praecox and vessel disease in white matter fasciculi.
schizophrenia. A common theme of these formulations is a In short, one could neatly summarise the pathophysiology of
disintegration or disruption of neuronal processing. From the many psychiatric and neurological conditions in terms of one or
perspective of the current narrative, these formulations came in two more forms of synaptopathy [108]. Synaptopathy is taken to mean
flavours. First, Wernicke’s sejunction hypothesis [103] emphasised any failure of synaptic function due to a variety pathological
A simplified scheme
Fig. 3 The left panel depicts interactions between (superficial and deep) pyramidal cells with inhibitory interneurons. Inhibitory
interneurons are partitioned into three subtypes (Parvalbumin positive PV, somatostatin SST, and vasoactive intestinal peptide expressing
interneurons, VIP), based upon optogenetic studies [217]. This schematic illustrates the intimate relationship between pyramidal cells and
inhibitory interneurons responsible for excitation-inhibition balance and ensuing excitability in canonical microcircuits. Here, it is assumed
that PV interneurons are reciprocally connected to the pyramidal cells through perisomatic compartments, while SST cells form synapses on
their dendrites. The right panel shows a simplified architecture used in many modelling studies. Here, the recurrent inhibitory (PV/pyramidal
cell) dynamics have been absorbed into an inhibitory recurrent connection, while the SST/VIP interneurons provide (dendritic) inhibitory
drive. This allows one to map models of Interneuron Network Gamma (ING) oscillations onto canonical microcircuits used in dynamic causal
modeling (see next figure). In this setting, the Pyramidal ING (PING) model emphasizes recurrent interactions among PV cells, as modeled by
the inhibitory recurrent connections on superficial pyramidal cells. In contrast, the ING model corresponds to the influence of (SST/VIP)
inhibitory interneurons on pyramidal cells. Please see [81] for further details of how these simplified architectures are used to model
attentional effects on evoked responses.
mechanisms (i.e., formation, structure, metabolism, etc.). For disorders may be attributable to aberrant precision control, which
example, the severe loss of synapse density in human Alzheimer’s inherits from synaptopathies that confound neuromodulation.
disease, frontotemporal dementia, Parkinson’s disease, Progressive This narrative first emerged in theoretical treatments of halluci-
Supranuclear Palsy, etc. [109]. nosis in synucleinopathies [126–128] and, from a computational
But what kind of synaptopathy? A coarse-grained overview of the perspective, was developed in schizophrenia [129–133] and
synaptic theories of schizophrenia suggests that the synaptopathy autism research [134–136]. Since that time, it has become difficult
in question is of a neuromodulatory sort; implicating classical to find a psychiatric condition that has not been considered
neuromodulatory (ascending) neurotransmitter systems [110–113], through the lens of aberrant precision, in one form or another
GABAergic neurotransmission and NMDA receptors [113–116], [21, 135, 137–161].
where synchronous interactions between fast spiking inhibitory A crosscutting theme in many of these accounts is a putative
interneurons and pyramidal cells may set the excitability of failure of sensory attenuation [53, 93–95, 98, 100, 159, 162–165]. In
canonical microcircuits [88, 91, 116–119]. The same story emerges other words, a neuromodulatory failure to attenuate the
from the functional genomics of schizophrenia, which point to the postsynaptic gain of neuronal populations reporting the con-
mechanisms that underwrite neuromodulation, synaptic gain sequence of action (in the motor domain or mediated by
control and ensuing excitation-inhibition balance in neuronal autonomic reflexes). A failure of sensory attenuation implies an
circuits [120, 121]. For example, a recent genome-wide association imbalance between the precision afforded sensory prediction
study [122] of 76,755 individuals with schizophrenia highlights the errors and the precision of prediction errors deeper in neuronal
importance of GRIN2A (which encodes a subunit of the NMDA hierarchies that mediate or maintain prior beliefs of a subpersonal
receptor), while a recent meta-analysis of whole exomes [123] or propositional nature. The ensuing imbalance is often read as a
emphasises the role of GRIN2A, as well as GRIA3 (which encodes a loss of precision or confidence in prior beliefs, relative to the
subunit of the AMPA receptor). Finally, it is worth noting that nearly sensory evidence at hand. In autism, this imbalance is consistent
all psychoactive drugs, including psychedelics, target classical with a failure of sensory attenuation and an inability to ignore the
neuromodulatory or NMDA receptors: e.g., [124, 125]. sensorium, which may or may not be associated with the
In short, many psychiatric, neurological and neurodegenerative neuromodulatory role of oxytocin [135, 166–170].
disorders could be described as arising from pernicious synapto- In schizophrenia, the story is a little more involved; in the sense
pathies that lead to a functional disintegration of neuronal that a failure of sensory attenuation provides an apt explanation
message passing in cortical and subcortical hierarchies. The for resistance to illusions (that normally depend upon precise prior
particular synaptopathies in question implicate neuromodulatory beliefs) and failures to elicit mismatch oddball responses (because
gain control, of the sort required to deploy precision (i.e., selective everything is surprising). Some people then interpret delusional
attention and sensory attenuation) during inference and planning ideation as the brain’s attempt to make sense of unattenuated
in the Bayesian brain. prediction errors: see also [152, 155, 171, 172]. However,
hallucinatory phenomena require a slightly more delicate argu-
ment; usually along the lines of a compensatory increase in prior
TOO MUCH OR TOO LITTLE? precision that could manifest as a form of paradoxical lesion. In
This completes the next step in the narrative; namely, the other words, in attempt to override precise sensory prediction
psychopathology (i.e., false inference) characteristic of psychiatric errors, higher levels learn to ignore sensory evidence and—
schizophrenia - predictable
schizophrenia - unpredictable
-0.5
-1
-1.5
-2
V1 R V5 L V5 R IT L IT R PC L PC
cortical source
Fig. 4 Synaptic gain control in schizophrenia. This figure summarizes an early example of using DCM of event related responses (ERPs), to
assess the modulation of synaptic efficacy in schizophrenia using EEG [191]. ERPs were elicited in a visual target detection paradigm, under
predictable and unpredictable conditions, in neurotypical and schizophrenic subjects, respectively. Sources—modeled as small cortical
patches (lower left panel)—include: a midline visual source (V1), right and left sources near the temporoparietal junction (V5), right and left
inferotemporal sources (IT) and bilateral superior parietal sources (PC). The distributed network connecting these sources (upper right panel)
entails top-down connections from PC and IT to V5 that send backward connections to V1 (black lines); reciprocal forward connections (red
lines); and intrinsic connections for each source (black loops). The principal components of predicted and observed ERPs (upper right panel)
show a pronounced difference in the evoked responses of normal subjects to predictable and unpredictable targets around 300 to 400 ms
after stimulus onset (compare the red and blue traces). This difference is attenuated in the schizophrenia patients (green and magenta). The
lower right panel reports (log scaling of ) intrinsic connections and their 90% posterior confidence intervals, for predictable versus
unpredictable targets, for patients (white) and controls (teal). The notable thing here is a failure of predictability to modulate the intrinsic gain
(i.e., self-connectivity) throughout the hierarchy in the schizophrenic subjects.
GABAergic modulation of deep pyramidal populations and different disorders. This could be regarded as one kind of
inhibitory interneurons. The authors then applied this model to precision medicine [200]; however, there is another kind that
patients with frontotemporal lobar degeneration, showing that speaks to characterising each individual in the sense of
“the phasic inhibition of deep cortico-cortical pyramidal neurons personalised medicine. Computational approaches in this setting
following tiagabine, but not placebo, was a function of GABA are potentially important in stratifying and classifying various
concentration” [199]. clinical phenotypes for the mechanistic studies outlined above.
Similar approaches have been considered in Parkinson’s disease This usually involves summarising a subject’s behaviour in terms
and schizophrenia, spanning cell cultures to human subjects. The of a computational model of that behaviour. Traditionally, this has
basic idea behind these proposals is to link scales through been a descriptive (e.g., reinforcement learning) model. This
Bayesian model comparison. In other words, optimise the model represents a weak kind of phenotyping because there are no well-
of neuronal microcircuits in ex vivo cell cultures, in terms of developed process theories for reinforcement learning. A strong
synaptic time constants of intrinsic connections, using Bayesian complement to computational phenotyping can however be
model inversion and selection. One can then use the posterior formalised using something called the complete class theorem
estimates as priors for the next scale (e.g., small animal [201, 202]. This basically says that there for any pair of behaviours
preparations, through to humans). Crucially, the superordinate and reward functions, there exists some priors that render the
scale introduces additional parameters (e.g., extrinsic connections behaviour Bayes optimal. This could be read as licensing any
among cortical areas) that are estimated more efficiently, having Bayesian ‘just so’ stories about choice behaviour [203]. However, it
resolved uncertainty about the parameters that are shared with has a deeper and more pragmatic implication.
lower scales. In short, by recursively using yesterday’s (one scale) It means that any given patient can be fully characterised by her
posteriors as tomorrow’s (next scale) priors, one has a seamless prior beliefs under ideal (active Bayesian inference) observer
and principled approach to integrating across scales. assumptions. This motivates the difficult game of inferring what
generative model this subject is using, based purely upon their
behavioural or neuronal responses [204–206]. Although at an early
COMPUTATIONAL PHENOTYPING AND NOSOLOGY stage, this sort of computational phenotyping shows promise, in
The narrative so far has moved from the theoretical frameworks, the sense that it can be more efficient than simply trying to
within which to place synaptopathy, to the requisite modelling of classify or stratify patients on the basis of their responses per se
empirical data to identify the synaptopathy that characterises [20]. This approach, sometimes called generative embedding,
Fig. 5 Measuring synaptopathy in vivo. This figure summarises the results of a DCM study of seizure activity in mice—induced with PTZ—in
NMDA receptor antibody (Ab) negative and positive cohorts, respectively [197]. Its agenda was to examine the effects of PTZ and receptor
antibodies on synaptic efficacy; summarised here in terms of the principal [eigen] component of changes in intrinsic connectivity. The
principal component of time constants implicated superficial pyramidal and spiny stellate cell changes (A), while the principal component of
connectivity-strength reflects changes in the coupling of spiny stellate to superficial pyramidal cells (B). The two principal components (of
variations in time constants and intrinsic connectivity) constitute a parameter space summarising the modulation of synaptic efficacy. For
each point in this parameter space, one can simulate the spectral responses one would observe data space. C in this panel, estimates of
synaptic efficacy for experimental subjects in the four conditions (pre-and post-PTZ, with and without antibodies) are shown as coloured dots.
These are superimposed on the predicted log mean delta power (with selected centile isoclines). The basic message here is that seizure
induction with PTZ shifts synaptic efficacy into regimes of high delta (c.f., slow-wave activity), and that this effect is more marked in the
presence of NMDA receptor antibodies. dp, deep pyramidal cells; ii, inhibitory interneurons; sp, superficial pyramidal cells; ss, spiny stellate
cells. Adapted with permission from the authors from.
has been applied both to the generative models a subject might modelling [210]. In principle, finding the right generative model
use to specify her behaviour and the generative models the for a patient guarantees the best predictions. This is because a
researcher uses to explain her physiological or neuronal responses model with the greatest evidence precludes overfitting and
[4, 207]. ensures generalisability [211]; namely, the generalisation from old
A related application of generative models to phenotyping (i.e., what has happened to the patient in the past) to new data
involves constructing dynamic causal models of hidden variables (what will happen to her in the future).
or states that underwrite the signs and symptoms expressed by
patients over weeks and years. This was a notable outcome of an
Ernst Strüngmann forum on computational psychiatry [208], COMPUTATIONAL NEUROPSYCHOLOGY
which concluded that conventional diagnoses—based upon A final part of the narrative is that a generative model of
traditional nosology—could play an important role in this kind pathophysiology, and ensuing psychopathology, allows one to
of computational nosology. However, the role was not to stratify perform in silico or synthetic experiments [212–214]. These entail
patients but rather as data features that supply evidence for or optimising the parameters of a model of a particular subject or
against models of the mapping between pathophysiology and cohort, and seeing what would happen if one increased or
psychopathology that are hidden from direct observation. decreased certain synaptic efficacies, e.g., [197]. Alternatively, one
A key tenet of this approach is to cast pathophysiology and can emulate extremes of aberrant precision by simply deleting
psychopathology as slowly fluctuating dynamical processes. This connections in active inference models of diagnostic paradigms,
speaks to a separation of temporal scales between processes like e.g., [215]. In these computational studies, the extrinsic (between
active inference [206], active learning [32] and structure learning node) connections are usually assigned to likelihood mappings
[209]. The mandatory coupling among these (fast and slow) relating sensory observations to representations or expectations
processes may underwrite cyclothymic disorders, relapsing- about hidden states of affairs. Conversely, intrinsic (within node)
remitting presentations and neurodevelopmental determinants. connections are usually treated as embodying prior beliefs about
An illustrative example of modelling the dynamics of schizoaffec- state transitions and narratives that characterise the paradigm in
tive disorders can be found in [208]. question.
An important aspect—of this application of computational One obvious advantage of being able to create a ‘digital twin’ of
psychiatry—is the quantification of uncertainty about how a a patient in silico, is the ability to perform synthetic lesion studies
patient will behave, and what is likely to happen to her in the and psychopharmacology, to test various hypotheses about the
future. Indeed, uncertainty quantification is a primary focus of effect of therapeutic interventions. This is probably best illustrated
related approaches in quantitative epidemiology and meteorol- in the setting of epilepsy, particularly, in predicting the effects of
ogy; especially in the context of forecasting and scenario pharmacological and surgical interventions; e.g., [216].